Prosody, intelligibility and familiarity in speech perception

Abstract

This thesis concerns the relationship between speech intelligibility and speech prosody, and the role that speech prosody plays in the perceptual advantage that occurs when listening to a familiar talker. A parallel corpus of 90 sentences (each spoken by three talkers of varying intelligibility) was used to create resynthesized stimuli in which fundamental frequency (f0), intensity, and patterns of syllable duration were swapped between all possible pairs of talkers. An additional 90 sentences were reserved for use as training stimuli. Findings from speech-in-noise tasks suggest that the contribution of prosody to intelligibility varies considerably across talkers, evidenced by differences in sentence intelligibility after prosodic replacement from different prosodic donors. In particular, high-intelligibility talkers need not have particularly "good" or "helpful" prosody if their intelligibility rests on articulatory strategies that emphasize robust segmental cues, while talkers with relatively good prosody may nonetheless have low intelligibility (due to non-prosodic factors). Post-hoc acoustic analyses of the stimuli (interpreted in light of the behavioral results) suggest that many acoustic measures index both prosodic and non-prosodic aspects of the speech signal, whereas acoustic measures that reflect only the prosodic component of intelligibility are harder to find. Mean f0 range and mean f0 dynamicity appear to be the most promising measures in this regard, but utterance-final creaky voicing presents a challenge to interpretation due to its exaggeration of f0-related measures. Results of the familiarization experiments were inconclusive; listeners trained on different talkers showed different degrees of task adaptation during familiarization/training, but were unable to generalize to a testing phase involving multiple talkers presented in random order. Thus the contribution of prosody to the familiar talker advantage remains unclear.